Perspicere at SemEval-2026 Task 2: Modeling Longitudinal Valence and Arousal via Dense Embeddings and Agentic Reasoning
Year: 2026
Type: conference-paper
Abstract: This paper presents our system for SemEval 2026 Task 2 (Subtask 1), modeling affect assessment as a longitudinal trajectory.We evaluate a tripartite affective framework of escalating contextual complexity, spanning zero-context feature extraction, latent temporal modeling via LSTM, and explicit semantic reasoning via the Teacher-Guided Clinical Reasoning Agent utilizing in-context learning.Our results show that robust static extraction outperforms explicit sequence modeling.Specifically, Matryoshkadistilled embeddings (Jasper) paired with XG-Boost provided the best balance of speed and accuracy when utilizing the full training corpus (Valence composite r = 0.654, a 17.4% improvement compared with the baseline), mitigating the severe overfitting observed on partitions of the dataset.Additionally, we uncover a distinct agentic advantage: although the reasoning agent trailed mathematical regressors in tracking high-frequency fluctuations, its SOTA psychological profiling yielded the highest Between-User Valence correlation (r = 0.725), demonstrating its efficacy in userlevel affective profiling.Finally, a persistent "arousal bottleneck" confirms the limitations of text-only modeling for physiological activation.
Language: English
Topic: Emotion and Mood Recognition
Subfield: Experimental and Cognitive Psychology
Field: Psychology
Domain: Social Sciences
Open Access status: hybrid
OpenAlex ID: W7166816339